Server engine · Rust

One semantic memory for everything your team reads, hears and sees.

Mnemoria ingests text, email, documents, audio and video frames into a single store — then derives embeddings, reranks, and serves retrieval. Built multi-tenant and access-controlled from the start, accelerated on Metal and CUDA, and runnable on a zero-dollar stack with no managed-SaaS lock-in.

Capabilities

Retrieval that scales piece by piece.

Multi-modal in one store

Text, email, documents, audio (STT) and visual (OCR + embedding) ingestion and retrieval — unified in a single index.

Detachable cube architecture

Embedding, reranking, STT and OCR run as independent, hot-swappable serving units with their own admission control. Scale a bottleneck without touching the rest.

Local & accelerated

Metal on Apple, CUDA on NVIDIA. Qwen3-Embedding-0.6B and Qwen3-Reranker-0.6B run where your hardware is.

Streaming transcription

Real-time SSE transcription keeps audio search responsive — results stream as the model hears them.

Multi-tenant & access-controlled

Keycloak auth, RBAC and Postgres row-level security. Built for teams, with isolation by default.

Zero-dollar stack

Production-grade retrieval without managed-SaaS lock-in. Own the whole pipeline, top to bottom.

Run Mnemoria as your team’s memory.

Start with the docs, clone the repo, and stand up a multi-tenant retrieval engine on hardware you already own. Looking for the desktop app? That is the MnemoVi! Local Video Knowledge Base.

Running on this engine

The engine, and the tools that run on it

Mnemoria is the indexing layer: it turns your own media into something searchable, locally. These are what it looks like once it has an application around it.